136879 research outputs found
Sort by
Economic Indicators of the College Station - Bryan MSA, December 2024
No full textThe Business-Cycle Index increased 0.4% from September 2024 to October 2024. The local unemployment rate remained steady at 3.2% in October 2024 compared to September 2024. Local nonfarm employment increased by 0.2% from September 2024 to October 2024. Inflation-adjusted taxable sales increased by 3.1% from September 2024 to October 2024. Among comparable college towns, the percentage of 20-24 aged residents was highest in the College Station-Bryan MSA at 17%. The College Station-Bryan MSA had the highest percentage of population growth at 1.3% among the selected college towns
Generating Fluent Text through Curriculum Learning and Disfluency Augmentation
No full textThere are several voice-only software systems available, however, language models employed in these systems are typically trained on huge, fluent, written text corpora. As a result, there is a misalignment between how language models are trained and how they are employed in voice-only systems, as speech frequently contains many disfluencies.
Depending on the individual speaking and the context, such disfluencies can occur occasionally or frequently. Because of this structure of disfluencies, we developed a synthetic disfluency amplification process. This method produces progressive versions of the original text with varying degrees of disfluency.
We combine this with Curriculum Learning to develop a novel training paradigm for producing fluent text from disfluent text, drawing analogies to the human learning process. Curriculum learning optimally utilizes the structure of our generated synthetic disfluent data, as more basic samples (i.e. more fluent) are fed to the model before more complex samples (i.e. more disfluent). In contrast to the random data exposure paradigm, this focuses on a simple-to-complex learning process.
We comprehensively evaluate our results on the Switchboard Penn Treebank-3 test set by comparing them to state-of-the-art parsing-based and translation-based models using word-based statistics. Our model surpasses existing techniques in terms of word-based precision and has shown good word-based recall and F1 scores
